Discrete Laplace Mechanism

• Discrete Laplace Mechanism is a differential privacy method that perturbs discrete data with two-sided geometric noise for rigorous privacy guarantees.
• It is widely applied in secure multi-party computation, local, and distributed privacy protocols, benefiting from efficient noise sampling and parameter tuning.
• The mechanism supports both pure (ε,0)-DP and approximate (ε,δ)-DP settings, with extensions like truncation and infinite divisibility to optimize utility and scalability.

The discrete Laplace mechanism (sometimes called the geometric mechanism) is a core construction for achieving differential privacy (DP) in discrete-valued data settings. It is used in sensitive data analysis, secure multi-party computation (MPC), local differential privacy (LDP), and distributed protocols, owing to its theoretical guarantees, computational efficiency, and versatility for both pure (ε,0)-DP and approximate (ε,δ)-DP. The mechanism operates by perturbing query outputs via noise drawn from a two-sided geometric (discrete Laplace) distribution, with parameter selection determined by the sensitivity of the function and the desired privacy level.

1. Formal Definitions and Probability Mass Functions

The discrete Laplace mechanism is defined for integer-valued or grid-valued data. Let $x \in \mathbb{Z}^n$ be an input with $\ell_1$-sensitivity $\Delta$, and fix $\varepsilon > 0$ as the privacy budget.

• Classic Discrete Laplace (Geometric) Mechanism: For each output, independently sample $\eta_j \sim \mathrm{DLap}(p)$ with $p = e^{-\varepsilon / \Delta}$. The pmf is

$$\Pr[\eta = k] = \frac{1-p}{1+p} p^{|k|}, \quad k \in \mathbb{Z}.$$

The privatized output is $\tilde{x} = x + \eta$ (Hillebrand et al., 7 May 2026).

• Radius-Truncated Discrete Laplace: For constrained output domains or local DP with bounded outputs, define a truncation radius $r$ and restrict noise to $|k| \le r$:

$\ell_1$0

(Zamani et al., 10 May 2026).

• Bounded/Truncated Discrete Laplace for MPC: For $\ell_1$1, and “bounding” parameter $\ell_1$2, noise is supported on $\ell_1$3 as

$\ell_1$4

where $\ell_1$5 and $\ell_1$6 normalizes over the support (Tjuawinata et al., 10 Mar 2025).

These mechanisms can be tailored for local, central, or distributed DP settings using suitable parameterizations and output supports.

2. Differential Privacy Guarantees: Pure and Approximate

The classic discrete Laplace mechanism achieves pure $\ell_1$7-differential privacy under modest assumptions:

• Privacy Proof (Classic): For any two neighboring $\ell_1$8, the output distributions $\ell_1$9 satisfy

$\Delta$0

for any output $\Delta$1, by direct computation of the pmf ratio and exploiting sensitivity bounds (Hillebrand et al., 7 May 2026, Zamani et al., 10 May 2026).

• Truncated/Bouned Case (Zero-Failure): For bounded mechanisms, as long as $\Delta$2 (with $\Delta$3 the $\Delta$4-sensitivity or bounding parameter), all outputs in $\Delta$5 have strictly positive probability and privacy holds with no “failure” probability $\Delta$6 (Tjuawinata et al., 10 Mar 2025).
• Approximate $\Delta$7-LDP in Sparse/Truncated Case: For mechanisms restricting outputs to a small neighborhood (support size $\Delta$8), privacy defect $\Delta$9 admits a precise decomposition in terms of support leakage and overlap contributions; explicit formulas enable tight $\varepsilon > 0$0 characterization and parameter selection to guarantee target privacy for given support size (Zamani et al., 10 May 2026).

3. Mechanism Sampling, Computation, and Efficient Implementation

Efficient sampling is central to practical deployment:

• Exact Geometric Sampling: Draw sign $\varepsilon > 0$1 (probability $\varepsilon > 0$2 each), sample $\varepsilon > 0$3 from geometric$\varepsilon > 0$4: $\varepsilon > 0$5, then return $\varepsilon > 0$6.
• Strict Polynomial-Time Samplers: For histogram release, implementers use inverse transform sampling on precomputed CDFs with (n+1)-wise independence for reproducible randomness and computational efficiency; word size and arithmetic precision are chosen to guarantee strict polynomial-time sampling in the finite computational model (Balcer et al., 2017).
• Multi-Party Setting (MPC): Sampling is divided into (a) offline phase for noise generation using secret-sharing and (b) lightweight online phase for applying generated noise to the secret-shared statistic; this enables high-throughput, failure-free deployments (Tjuawinata et al., 10 Mar 2025).
• Fast Distributed Sampling: For infinitely divisible variants, negative binomial and geometric samplers are used in concert with efficient conditional allocation (e.g., via Dirichlet-multinomial sampling in the multi-scale discrete Laplace) for $\varepsilon > 0$7 per-party sample time even for high sensitivity or privacy budgets (Harrison et al., 7 Apr 2025).

4. Utility Analysis: Bias, Variance, Mean Squared Error

The utility of discrete Laplace mechanisms is characterized by bias, variance, and MSE as functions of mechanism parameters:

• Bias and Variance (Bounded Case):

$\varepsilon > 0$8

For $\varepsilon > 0$9, this yields $\eta_j \sim \mathrm{DLap}(p)$0 and bias exponentially small in $\eta_j \sim \mathrm{DLap}(p)$1 (Tjuawinata et al., 10 Mar 2025).

• Classic Discrete Laplace:

$\eta_j \sim \mathrm{DLap}(p)$2

This is suboptimal compared to continuous staircase mechanisms for large $\eta_j \sim \mathrm{DLap}(p)$3, but improvements are possible via multi-scale or generalized discrete Laplace constructions (Harrison et al., 7 Apr 2025).

• Multi-Scale and Generalized Variants: The multi-scale DLap achieves

$\eta_j \sim \mathrm{DLap}(p)$4

matching order-optimal bounds for pure DP additive noise (Harrison et al., 7 Apr 2025).

• Empirical and Application-Based Utility: Unbiased postprocessing yields drastic improvements for structure-preserving statistics (e.g., entropy, profile estimation), especially under distributed or federated settings (Hillebrand et al., 7 May 2026).

5. Extensions: Postprocessing, Truncation, and Infinitely Divisible Variants

• Postprocessing for Unbiased Estimation: Given a function $\eta_j \sim \mathrm{DLap}(p)$5 of the data, unbiased estimators $\eta_j \sim \mathrm{DLap}(p)$6 can be constructed via local corrections of the output using finite differencing; performance is bounded by the function’s local structure and problem dimension (Hillebrand et al., 7 May 2026).
• Matching Continuous or Staircase Mechanisms: Adding symmetric continuous randomness (scaled for cell-wise convolution) to DLap output yields exact Laplace or staircase noise, so the discrete Laplace is a universal primitive for all such mechanisms on integral data (Hillebrand et al., 7 May 2026).
• Truncated/Bouned/Compact Variants: Mechanisms such as truncated discrete Laplace provide minimal-distortion, failure-free output for given support sizes, with exact formulas to trade off distortion, privacy, and computational overhead (Zamani et al., 10 May 2026, Tjuawinata et al., 10 Mar 2025).
• Infinitely Divisible Mechanisms: The generalized discrete Laplace (GDL) and multi-scale discrete Laplace (MSDLap) mechanisms are designed for distributed implementations in which individual noise shares sum to an overall DP-protective noise. Infinite divisibility underpins scalability and robustness in asynchronous/distributed settings (Harrison et al., 7 Apr 2025).

Mechanism	DP Form	Error (MSE)	Infinite Divisibility	Truncation
DLap (classic)	ε-DP	$\eta_j \sim \mathrm{DLap}(p)$7	Yes	Optional
GDL	ε-DP	$\eta_j \sim \mathrm{DLap}(p)$8	Yes	No
MSDLap	ε-DP	$\eta_j \sim \mathrm{DLap}(p)$9	Yes	No
Bounded DLap (TDL)	ε-DP	$p = e^{-\varepsilon / \Delta}$0	Yes (if unconstrained)	Yes (by design)
Discrete Gaussian	(ε,δ)-DP	Slightly lower than DLap	No	Yes (costly in MPC)

6. Practical Applications and Performance Benchmarks

The discrete Laplace mechanism and its variants are fundamental to:

• Secure Multi-Party Computation (MPC): Provides practical, failure-free mechanisms with empirically validated throughput on platforms such as ABY, with online perturbation requiring only a few milliseconds per sample and circuit complexity comparable to or better than alternative DP constructions (Tjuawinata et al., 10 Mar 2025).
• Finite Precision Environments: Discrete mechanisms avoid pitfalls of floating-point arithmetic (such as bias and privacy loss due to rounding), enabling strict polynomial-time DP primitives amenable to analysis and reliable deployment (Balcer et al., 2017).
• Local DP and Histogram Release: Permits compact or sparse representations and computationally efficient estimation, with provable lower bounds for per-bin and simultaneous error (Balcer et al., 2017).
• Federated and Distributed Settings: Infinitely divisible noise mechanisms such as GDL and MSDLap enable scalable distributed DP aggregation protocols with order-optimal utility, including shuffle DP protocols improving on previous MSE and message complexity bounds (Harrison et al., 7 Apr 2025).
• Empirical Use Cases: Unbiased estimators and discrete Laplace postprocessing outperform naive estimators in entropy, profile, and graph-analytic tasks, demonstrating the practical significance of principled postprocessing and estimator design (Hillebrand et al., 7 May 2026).

7. Trade-Offs, Design Principles, and Theoretical Boundaries

The selection and tuning of discrete Laplace mechanisms inherently involve trade-offs among privacy, utility, computational complexity, and output sparsity:

• Privacy-Utility Tradeoff: Higher privacy (low ε) demands larger noise and increases variance; support truncation minimizes failure but may increase distortion unless parameters are chosen using explicit design formulas (Zamani et al., 10 May 2026).
• Support Size and Sparsity: In sparse/truncated settings, the support size $p = e^{-\varepsilon / \Delta}$1 must exceed a minimum threshold to satisfy (ε,δ)-DP, determined precisely by the domain diameter and privacy parameters (Zamani et al., 10 May 2026). For compact histograms, efficient representations restore optimal per-bin error while avoiding explicit enumeration (Balcer et al., 2017).
• Computational Constraints: Polynomial-time discrete mechanisms are preferred in finite computational models; infinitely divisible mechanisms are essential for distributed and asynchronous deployments (Harrison et al., 7 Apr 2025).
• Comparison with Discrete Gaussian: While discrete Gaussian noise achieves slightly lower variance under approximate DP, for pure DP and ease of implementation in MPC or distributed settings, discrete Laplace and its efficient, bounded variants match or exceed practical efficiency, with guaranteed zero failure (Tjuawinata et al., 10 Mar 2025).

Discrete Laplace mechanisms thus represent a theoretically complete and practically versatile toolkit for differential privacy in discrete, distributed, and resource-constrained environments.